Skip to content

vllm.models.kimi_k3.nvidia.specialized_decode.cutedsl.oproj.producer

O-projection producer for the split and fused AttnRes paths: 112 CTAs of 64 output rows. Each warp keeps rows 16-row weight tiles of one of 4 * rows K parts in registers (TP8 stages some K blocks in shared memory), runs BF16 MMA chains and the K parts are reduced in order. Weights load before the PDL wait; the split producer triggers its consumer right after it. With group_tiles the batch runs in groups of that many 8-token tiles: each group's x arrives through a double-buffered cp.async ring during the previous group's MMAs, and is reduced and published during the next one's. With twoshot the partial chunks go to their owners' receive slots instead (the first shot of common.twoshot; args 2-4: receive view, owners' bases, rank). OprojUmma is the tcgen05 engine of the same producer for B>8.

Classes:

Functions:

  • pack8 –

    Four BF16x2 words of values[0:8].

OprojProducer

Methods:

  • multiply –

    FP32 chains of tokens [first, first + group); x from xr (registers) or

  • publish –

    Sum the K parts of tokens [first, first + group) in order; multicast them.

  • schedule –

    (8-token tiles, tiles per group, groups) of batches tokens.

  • stage_x –

    One cp.async group: x tokens [first, first + group) into ring buffer.

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
class OprojProducer:
    def __init__(self, k, *, staged_blocks, rows=1, group_tiles=0, twoshot=False):
        self.k, self.staged, self.rows = k, staged_blocks, rows
        self.twoshot = twoshot
        self.parts = 4 * rows
        self.kblocks = k // (32 * self.parts)
        self.register_blocks = self.kblocks - self.staged
        self.group_tiles = group_tiles
        # Single pass: TP8 copies the batch to shared memory, TP16 loads registers.
        self.shared_x = self.staged > 0 and not group_tiles
        self.register_x = k == 768 and not group_tiles
        self.x_stride = k // 8 + 4
        self.chains = max(1, 4 // (rows * max(1, group_tiles)))

    def schedule(self, batches):
        """(8-token tiles, tiles per group, groups) of ``batches`` tokens."""
        tiles = (batches + 7) // 8
        per_group = self.group_tiles or tiles
        return tiles, per_group, (tiles + per_group - 1) // per_group

    @cute.jit
    def allocate(self, batches: cutlass.Constexpr[int]):
        nt, gt, ng = self.schedule(batches)
        smem = cutlass.utils.SmemAllocator()
        red = alloc(smem, Float32, (gt, 4, self.parts, 8, 20))
        rows = nt * 8 if self.shared_x else 0
        if cutlass.const_expr(self.group_tiles):
            rows = min(2, ng) * gt * 8
        xs = alloc(smem, Uint32, rows * self.x_stride * 4 if rows else 1)
        staged = alloc(smem, Uint32, max(1, 16 * self.staged * self.rows * 2 * 32 * 4))
        return red, xs, staged

    @cute.jit
    def load_weight(self, weight, staged, cta):
        tid, _, _ = cute.arch.thread_idx()
        lane, warp = tid % 32, tid // 32
        group, part = lane // 4, lane % 4
        tile = cta * 4 + warp // self.parts * self.rows
        kq = warp % self.parts
        base = (tile * 16 + group) * self.k + kq * self.kblocks * 32 + part * 8
        if cutlass.const_expr(self.staged > 0):
            for j in cutlass.range_constexpr(self.staged):
                for t in cutlass.range_constexpr(self.rows):
                    unit = (warp * self.staged + j) * self.rows + t
                    dst = staged.iterator + (unit * 2 * 32 + lane) * 4
                    src = weight.iterator + base + t * 16 * self.k
                    src += (self.register_blocks + j) * 32
                    stage_weight(dst, src)
                    stage_weight(dst + 128, src + 8 * self.k)
            cute.arch.cp_async_commit_group()
        wr = cute.make_rmem_tensor((self.register_blocks, self.rows, 2, 4), Uint32)
        for j in cutlass.range_constexpr(self.register_blocks):
            for t in cutlass.range_constexpr(self.rows):
                for r in cutlass.range_constexpr(2):
                    src = weight.iterator + base + (t * 2 + r) * 8 * self.k + j * 32
                    words = load_vector(src, cache="weight")
                    for v in cutlass.range_constexpr(4):
                        wr[j, t, r, v] = words[v]
        return wr

    @cute.jit
    def stage_x(self, x, xs, first: Constexpr[int], buffer: Constexpr[int]):
        """One cp.async group: x tokens [first, first + group) into ring ``buffer``."""
        tid, _, _ = cute.arch.thread_idx()
        chunks: Constexpr = self.k // 8
        count: Constexpr = min(x.shape[0] - first, self.group_tiles * 8)
        rows: Constexpr = buffer * self.group_tiles * 8
        for i in cutlass.range_constexpr((count * chunks + 511) // 512):
            chunk = tid + i * 512
            token, col = chunk // chunks, chunk % chunks
            dst = xs.iterator + ((rows + token) * self.x_stride + col) * 4
            if cutlass.const_expr((i + 1) * 512 <= count * chunks):
                copy_g2s(dst, at(x, first + token, col * 8))
            else:
                if chunk < count * chunks:  # noqa: SIM102
                    copy_g2s(dst, at(x, first + token, col * 8))
        cute.arch.cp_async_commit_group()

    def x_words(self, xs, row, kofs, j):
        return smem_words(xs.iterator + (row * self.x_stride + kofs // 8 + j * 4) * 4)

    @cute.jit
    def multiply(self, smem, wr, xr, first: Constexpr, buffer: Constexpr, M: Constexpr):
        """FP32 chains of tokens [first, first + group); x from ``xr`` (registers) or
        shared-memory rows (ring ``buffer``)."""
        red, xs, staged = smem
        _, gt, _ = self.schedule(M)
        tiles: Constexpr = min(gt, (M - first + 7) // 8)
        tid, _, _ = cute.arch.thread_idx()
        lane, warp = tid % 32, tid // 32
        group, part = lane // 4, lane % 4
        kofs = warp % self.parts * self.kblocks * 32 + part * 8
        acc = cute.make_rmem_tensor((gt, self.rows, self.chains, 4), Float32)
        acc.fill(0)
        for j in cutlass.range_constexpr(self.kblocks):
            wa, wb = [], []
            for t in cutlass.range_constexpr(self.rows):
                if cutlass.const_expr(j < self.register_blocks):
                    wa.append(tuple(wr[j, t, 0, v] for v in range(4)))
                    wb.append(tuple(wr[j, t, 1, v] for v in range(4)))
                else:
                    sj: Constexpr = j - self.register_blocks
                    unit = (warp * self.staged + sj) * self.rows + t
                    ptr = staged.iterator + (unit * 2 * 32 + lane) * 4
                    wa.append(smem_words(ptr))
                    wb.append(smem_words(ptr + 128))
            for nt in cutlass.range_constexpr(tiles):
                if cutlass.const_expr(self.register_x):
                    xv = tuple(xr[nt, j, v] for v in range(4))
                else:
                    row = buffer * gt * 8 + nt * 8
                    if cutlass.const_expr(self.group_tiles):
                        xv = self.x_words(xs, row + group, kofs, j)
                    else:
                        xv = (Uint32(0), Uint32(0), Uint32(0), Uint32(0))
                        if nt * 8 + group < M:
                            xv = self.x_words(xs, row + group, kofs, j)
                for t in cutlass.range_constexpr(self.rows):
                    for half in cutlass.range_constexpr(2):
                        chain: Constexpr = (j * 2 + half) % self.chains
                        a: tuple = (wa[t][2 * half], wb[t][2 * half])
                        a += (wa[t][2 * half + 1], wb[t][2 * half + 1])
                        b = (xv[2 * half], xv[2 * half + 1])
                        mma_into(acc[nt, t, chain, None], a, b, bf16=True)
        return acc

    @cute.jit
    def store_partials(self, red, acc, tiles: Constexpr):
        tid, _, _ = cute.arch.thread_idx()
        lane, warp = tid % 32, tid // 32
        group, part = lane // 4, lane % 4
        tile, kq = warp // self.parts * self.rows, warp % self.parts
        for nt in cutlass.range_constexpr(tiles):
            for t in cutlass.range_constexpr(self.rows):
                for v in cutlass.range_constexpr(4):
                    chains = [acc[nt, t, c, v] for c in range(self.chains)]
                    token, row = 2 * part + v % 2, group + (v // 2) * 8
                    red[nt, tile + t, kq, token, row] = tree_reduce(chains, add_rn)

    @cute.jit
    def publish(self, args, red, cta, first: Constexpr[int]):
        """Sum the K parts of tokens [first, first + group) in order; multicast them."""
        x, _, mailbox, mc, slot = args[:5]
        _, gt, _ = self.schedule(x.shape[0])
        count: Constexpr = min(x.shape[0] - first, gt * 8)
        tid, _, _ = cute.arch.thread_idx()
        if tid < 8 * count:
            reduced = cute.make_rmem_tensor(8, Float32)
            if cutlass.const_expr(self.group_tiles):
                # Token-major lanes: conflict-free 16-byte partial loads.
                token, out_half = tid % count, tid // count % 2
                row_tile = tid // count // 2
                tile, row = token // 8, token % 8
                for i in cutlass.range_constexpr(2):
                    column = out_half * 8 + i * 4
                    words = smem_words(at(red, tile, row_tile, 0, row, column))
                    sums = [bits_float(w) for w in words]
                    for q in cutlass.range_constexpr(1, self.parts):
                        words = smem_words(at(red, tile, row_tile, q, row, column))
                        sums = [add_rn(s, bits_float(w)) for s, w in zip(sums, words)]
                    for v in cutlass.range_constexpr(4):
                        reduced[i * 4 + v] = sums[v]
            else:
                out_half, row_tile, token = tid % 2, (tid // 2) % 4, tid // 8
                tile, row = token // 8, token % 8
                for v in cutlass.range_constexpr(8):
                    value = red[tile, row_tile, 0, row, out_half * 8 + v]
                    for q in cutlass.range_constexpr(1, self.parts):
                        part = red[tile, row_tile, q, row, out_half * 8 + v]
                        value = add_rn(value, part)
                    reduced[v] = value
            words = pack8(reduced, "latent")
            out_col = (cta * 4 + row_tile) * 16 + out_half * 8
            if cutlass.const_expr(self.twoshot):
                push(mailbox, mc, slot, first + token, out_col, words)
            else:
                offset = cute.crd2idx((slot, first + token, out_col), mailbox.layout)
                multicast_words(mc + Int64(offset) * 2, words)

    @cute.jit
    def compute(self, args, smem, wr, cta, trigger_after_wait: cutlass.Constexpr[bool]):
        x = args[0]
        red, xs, staged = smem
        M: Constexpr = x.shape[0]
        NT, GT, NG = self.schedule(M)
        tid, _, _ = cute.arch.thread_idx()
        lane, warp = tid % 32, tid // 32
        group, part = lane // 4, lane % 4
        kofs = warp % self.parts * self.kblocks * 32 + part * 8
        cute.arch.griddepcontrol_wait()
        if cutlass.const_expr(trigger_after_wait):
            cute.arch.griddepcontrol_launch_dependents()
        xr = cute.make_rmem_tensor((NT, self.kblocks, 4), Uint32)
        if cutlass.const_expr(self.group_tiles):
            self.stage_x(x, xs, 0, 0)
        if cutlass.const_expr(self.shared_x):
            for chunk in range(tid, M * (self.k // 8), 512):
                token, col = chunk // (self.k // 8), chunk % (self.k // 8)
                words = load_vector(at(x, token, col * 8))
                store_shared_words(
                    xs.iterator + (token * self.x_stride + col) * 4, words
                )
            cute.arch.sync_threads()
        if cutlass.const_expr(self.register_x):
            xr.fill(0)
            for nt in cutlass.range_constexpr(NT):
                token = nt * 8 + group
                if token < M:
                    for j in cutlass.range_constexpr(self.kblocks):
                        words = load_vector(at(x, token, kofs + j * 32))
                        for v in cutlass.range_constexpr(4):
                            xr[nt, j, v] = words[v]
        if cutlass.const_expr(self.staged > 0 and not self.group_tiles):
            cute.arch.cp_async_wait_group(0)
        for g in cutlass.range_constexpr(NG):
            if cutlass.const_expr(self.group_tiles):
                # x(g) and group g - 1's partials are visible: stage x(g + 1), publish.
                cute.arch.cp_async_wait_group(0)
                cute.arch.sync_threads()
                if cutlass.const_expr(g + 1 < NG):
                    self.stage_x(x, xs, (g + 1) * GT * 8, (g + 1) % 2)
                if cutlass.const_expr(g > 0):
                    self.publish(args, red, cta, (g - 1) * GT * 8)
            acc = self.multiply(smem, wr, xr, g * GT * 8, g % 2, M)
            if cutlass.const_expr(self.group_tiles):
                cute.arch.sync_threads()
            self.store_partials(red, acc, min(GT, NT - g * GT))
        cute.arch.sync_threads()
        self.publish(args, red, cta, (NG - 1) * GT * 8)

    @cute.jit
    def __call__(self, *args):
        launch(self.kernel, args, 112, 512, min_blocks_per_mp=1)

    @cute.kernel
    def kernel(self, args):
        cta, _, _ = cute.arch.block_idx()
        smem = self.allocate(args[0].shape[0])
        wr = self.load_weight(args[1], smem[2], cta)
        self.compute(args, smem, wr, cta, True)

multiply(smem, wr, xr, first, buffer, M)

FP32 chains of tokens [first, first + group); x from xr (registers) or shared-memory rows (ring buffer).

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
@cute.jit
def multiply(self, smem, wr, xr, first: Constexpr, buffer: Constexpr, M: Constexpr):
    """FP32 chains of tokens [first, first + group); x from ``xr`` (registers) or
    shared-memory rows (ring ``buffer``)."""
    red, xs, staged = smem
    _, gt, _ = self.schedule(M)
    tiles: Constexpr = min(gt, (M - first + 7) // 8)
    tid, _, _ = cute.arch.thread_idx()
    lane, warp = tid % 32, tid // 32
    group, part = lane // 4, lane % 4
    kofs = warp % self.parts * self.kblocks * 32 + part * 8
    acc = cute.make_rmem_tensor((gt, self.rows, self.chains, 4), Float32)
    acc.fill(0)
    for j in cutlass.range_constexpr(self.kblocks):
        wa, wb = [], []
        for t in cutlass.range_constexpr(self.rows):
            if cutlass.const_expr(j < self.register_blocks):
                wa.append(tuple(wr[j, t, 0, v] for v in range(4)))
                wb.append(tuple(wr[j, t, 1, v] for v in range(4)))
            else:
                sj: Constexpr = j - self.register_blocks
                unit = (warp * self.staged + sj) * self.rows + t
                ptr = staged.iterator + (unit * 2 * 32 + lane) * 4
                wa.append(smem_words(ptr))
                wb.append(smem_words(ptr + 128))
        for nt in cutlass.range_constexpr(tiles):
            if cutlass.const_expr(self.register_x):
                xv = tuple(xr[nt, j, v] for v in range(4))
            else:
                row = buffer * gt * 8 + nt * 8
                if cutlass.const_expr(self.group_tiles):
                    xv = self.x_words(xs, row + group, kofs, j)
                else:
                    xv = (Uint32(0), Uint32(0), Uint32(0), Uint32(0))
                    if nt * 8 + group < M:
                        xv = self.x_words(xs, row + group, kofs, j)
            for t in cutlass.range_constexpr(self.rows):
                for half in cutlass.range_constexpr(2):
                    chain: Constexpr = (j * 2 + half) % self.chains
                    a: tuple = (wa[t][2 * half], wb[t][2 * half])
                    a += (wa[t][2 * half + 1], wb[t][2 * half + 1])
                    b = (xv[2 * half], xv[2 * half + 1])
                    mma_into(acc[nt, t, chain, None], a, b, bf16=True)
    return acc

publish(args, red, cta, first)

Sum the K parts of tokens [first, first + group) in order; multicast them.

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
@cute.jit
def publish(self, args, red, cta, first: Constexpr[int]):
    """Sum the K parts of tokens [first, first + group) in order; multicast them."""
    x, _, mailbox, mc, slot = args[:5]
    _, gt, _ = self.schedule(x.shape[0])
    count: Constexpr = min(x.shape[0] - first, gt * 8)
    tid, _, _ = cute.arch.thread_idx()
    if tid < 8 * count:
        reduced = cute.make_rmem_tensor(8, Float32)
        if cutlass.const_expr(self.group_tiles):
            # Token-major lanes: conflict-free 16-byte partial loads.
            token, out_half = tid % count, tid // count % 2
            row_tile = tid // count // 2
            tile, row = token // 8, token % 8
            for i in cutlass.range_constexpr(2):
                column = out_half * 8 + i * 4
                words = smem_words(at(red, tile, row_tile, 0, row, column))
                sums = [bits_float(w) for w in words]
                for q in cutlass.range_constexpr(1, self.parts):
                    words = smem_words(at(red, tile, row_tile, q, row, column))
                    sums = [add_rn(s, bits_float(w)) for s, w in zip(sums, words)]
                for v in cutlass.range_constexpr(4):
                    reduced[i * 4 + v] = sums[v]
        else:
            out_half, row_tile, token = tid % 2, (tid // 2) % 4, tid // 8
            tile, row = token // 8, token % 8
            for v in cutlass.range_constexpr(8):
                value = red[tile, row_tile, 0, row, out_half * 8 + v]
                for q in cutlass.range_constexpr(1, self.parts):
                    part = red[tile, row_tile, q, row, out_half * 8 + v]
                    value = add_rn(value, part)
                reduced[v] = value
        words = pack8(reduced, "latent")
        out_col = (cta * 4 + row_tile) * 16 + out_half * 8
        if cutlass.const_expr(self.twoshot):
            push(mailbox, mc, slot, first + token, out_col, words)
        else:
            offset = cute.crd2idx((slot, first + token, out_col), mailbox.layout)
            multicast_words(mc + Int64(offset) * 2, words)

schedule(batches)

(8-token tiles, tiles per group, groups) of batches tokens.

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
76
77
78
79
80
def schedule(self, batches):
    """(8-token tiles, tiles per group, groups) of ``batches`` tokens."""
    tiles = (batches + 7) // 8
    per_group = self.group_tiles or tiles
    return tiles, per_group, (tiles + per_group - 1) // per_group

stage_x(x, xs, first, buffer)

One cp.async group: x tokens [first, first + group) into ring buffer.

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
@cute.jit
def stage_x(self, x, xs, first: Constexpr[int], buffer: Constexpr[int]):
    """One cp.async group: x tokens [first, first + group) into ring ``buffer``."""
    tid, _, _ = cute.arch.thread_idx()
    chunks: Constexpr = self.k // 8
    count: Constexpr = min(x.shape[0] - first, self.group_tiles * 8)
    rows: Constexpr = buffer * self.group_tiles * 8
    for i in cutlass.range_constexpr((count * chunks + 511) // 512):
        chunk = tid + i * 512
        token, col = chunk // chunks, chunk % chunks
        dst = xs.iterator + ((rows + token) * self.x_stride + col) * 4
        if cutlass.const_expr((i + 1) * 512 <= count * chunks):
            copy_g2s(dst, at(x, first + token, col * 8))
        else:
            if chunk < count * chunks:  # noqa: SIM102
                copy_g2s(dst, at(x, first + token, col * 8))
    cute.arch.cp_async_commit_group()

OprojUmma

Bases: OprojProducer

B>8 producer on tcgen05: 112 CTAs x 64 rows, tokens on N. Four loader warps stream 128B-swizzled W [64, 64] and x [N, 64] K blocks through a cp.async ring (W of the first stages before the PDL wait, the rest prefetched to L2), the MMA warp accumulates D [64, N] FP32 in TMEM with M=64 kind::f16 MMAs, and the loader warps round D to BF16 and publish it like OprojProducer.

Methods:

  • epilogue –

    D rows 16 * warp + lane (lane < 16: M=64 TMEM layout) -> FP32 [token, row]

  • geometry –

    (N, ring stage bytes, stages, TMEM columns).

  • load –

    cp.async rows x 64 BF16 of source from (row0, column) into a K-major

  • multiply –

    The MMA warp: wait for each K block, issue its four K16 MMAs from the

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
class OprojUmma(OprojProducer):
    """B>8 producer on tcgen05: 112 CTAs x 64 rows, tokens on N. Four loader warps
    stream 128B-swizzled W [64, 64] and x [N, 64] K blocks through a cp.async ring
    (W of the first stages before the PDL wait, the rest prefetched to L2), the MMA
    warp accumulates D [64, N] FP32 in TMEM with M=64 kind::f16 MMAs, and the loader
    warps round D to BF16 and publish it like ``OprojProducer``."""

    def __init__(self, k, *, twoshot=False, budget=184 * 1024):
        super().__init__(k, staged_blocks=0, twoshot=twoshot)
        self.blocks, self.budget = k // 64, budget

    def geometry(self, batches):
        """(N, ring stage bytes, stages, TMEM columns)."""
        n = (batches + 7) // 8 * 8
        stage = 8192 + n * 128
        stages = min(self.blocks, self.budget // stage)
        return n, stage, stages, max(32, 1 << (n - 1).bit_length())

    @cute.jit
    def load(self, source, ring, rows: Constexpr, valid: Constexpr, row0, column):
        """cp.async ``rows`` x 64 BF16 of ``source`` from (row0, column) into a K-major
        128B-swizzled tile; rows >= ``valid`` are left alone. ``valid`` < 0: weights
        (evict-first, every row)."""
        tid, _, _ = cute.arch.thread_idx()
        for i in cutlass.range_constexpr((rows * 8 + 127) // 128):
            c = tid + 128 * i
            row, chunk = c // 8, c % 8
            dst = ring + row // 8 * 1024 + row % 8 * 128 + (chunk ^ row % 8) * 16
            src = at(source, row0 + row, column + chunk * 8)
            if cutlass.const_expr(valid < 0):
                stage_weight(dst, src)
            elif cutlass.const_expr((i + 1) * 128 <= valid * 8):
                copy_g2s(dst, src)
            else:
                if c < valid * 8:  # noqa: SIM102
                    copy_g2s(dst, src)

    @cute.jit
    def multiply(self, ring, bars, tmem, n: Constexpr, stage: Constexpr):
        """The MMA warp: wait for each K block, issue its four K16 MMAs from the
        elected lane and release its stage (the last block: the accumulator)."""
        stages: Constexpr = self.geometry(n)[2]
        # kind::f16 descriptor: FP32 D, BF16 A and B, both K-major, N, M = 64
        instruction: Constexpr = 1 << 4 | 1 << 7 | 1 << 10 | n >> 3 << 17 | 4 << 24
        for kb in cutlass.range_constexpr(self.blocks):
            s: Constexpr = kb % stages
            wait(bars.iterator + s, kb // stages % 2)
            cute.arch.fence_proxy("async.shared", space="cta")
            fence_tmem()
            a = smem_desc(ring.iterator + s * stage, True)
            b = smem_desc(ring.iterator + s * stage + 8192, True)
            last: Constexpr = kb + 1 == self.blocks
            release: Constexpr = last or kb + stages < self.blocks
            barrier = bars.iterator + (2 * stages if last else stages + s)
            mma_bf16_block(tmem, a, b, instruction, int(kb > 0), barrier, int(release))

    @cute.jit
    def epilogue(self, args, ring, tmem, cta, n: Constexpr):
        """D rows 16 * warp + lane (lane < 16: M=64 TMEM layout) -> FP32 [token, row]
        tile in the drained ring -> 8-row BF16 chunks to the mailbox / partials."""
        x, _, mailbox, mc, slot = args[:5]
        M: Constexpr = x.shape[0]
        tid, warp, lane, _ = thread()
        values = load_tmem(tmem + (32 * warp << 16), self.geometry(M)[3])
        tile = view(ring.iterator, 0, Float32, (n, 68), (68, 1))  # [token, row]
        if lane < 16:
            for t in cutlass.range_constexpr(M):
                tile[t, 16 * warp + lane] = values[t]
        named_barrier(1, 128)
        for i in cutlass.range_constexpr((M * 8 + 127) // 128):
            c = tid + 128 * i
            if c < M * 8:
                token, part = c // 8, c % 8
                reduced = cute.make_rmem_tensor(8, Float32)
                for h in cutlass.range_constexpr(2):
                    words = smem_words(at(tile, token, part * 8 + h * 4))
                    for v in cutlass.range_constexpr(4):
                        reduced[h * 4 + v] = bits_float(words[v])
                words = pack8(reduced, "latent")
                column = cta * 64 + part * 8
                if cutlass.const_expr(self.twoshot):
                    push(mailbox, mc, slot, token, column, words)
                else:
                    offset = cute.crd2idx((slot, token, column), mailbox.layout)
                    multicast_words(mc + Int64(offset) * 2, words)

    @cute.jit
    def __call__(self, *args):
        launch(self.kernel, args, 112, 160, min_blocks_per_mp=1)

    @cute.kernel
    def kernel(self, args):
        x, weight = args[0], args[1]
        n, stage, stages, columns = self.geometry(x.shape[0])
        tid, warp, _, cta = thread()
        smem = cutlass.utils.SmemAllocator()
        ring = alloc(smem, Uint8, stages * stage, 1024)
        bars = alloc(smem, Uint64, 2 * stages + 1, 8)
        tmem_slot = alloc(smem, Uint32, 1)
        if tid == 0:
            for s in cutlass.range_constexpr(stages):
                cute.arch.mbarrier_init(bars.iterator + s, 256)
                cute.arch.mbarrier_init(bars.iterator + stages + s, 1)
            cute.arch.mbarrier_init(bars.iterator + 2 * stages, 1)
            cute.arch.mbarrier_init_fence()
        if warp == 4:
            cute.arch.alloc_tmem(columns, tmem_slot.iterator)
            cute.arch.relinquish_tmem_alloc_permit()
        fence_tmem(before=True)
        cute.arch.sync_threads()
        fence_tmem()
        tmem = tmem_slot[0]
        row0 = cta * 64
        if warp < 4:  # weights do not depend on the previous kernel
            if cutlass.const_expr(stages < self.blocks) and tid < 64:
                rest = at(weight, row0 + tid, stages * 64)
                prefetch_bytes(rest, (self.blocks - stages) * 128)
            for s in cutlass.range_constexpr(stages):
                self.load(weight, ring.iterator + s * stage, 64, -1, row0, s * 64)
                cute.arch.cp_async_mbarrier_arrive_noinc(bars.iterator + s)
        cute.arch.griddepcontrol_wait()
        cute.arch.griddepcontrol_launch_dependents()
        if warp < 4:
            for kb in cutlass.range_constexpr(self.blocks):
                slot: Constexpr = kb % stages
                tile = ring.iterator + slot * stage
                if cutlass.const_expr(kb >= stages):
                    wait(bars.iterator + stages + slot, (kb // stages - 1) % 2)
                    self.load(weight, tile, 64, -1, row0, kb * 64)
                    cute.arch.cp_async_mbarrier_arrive_noinc(bars.iterator + slot)
                self.load(x, tile + 8192, n, x.shape[0], 0, kb * 64)
                cute.arch.cp_async_mbarrier_arrive_noinc(bars.iterator + slot)
            wait(bars.iterator + 2 * stages, 0)
            fence_tmem()
            # The MMAs (async proxy) read the ring the epilogue now overwrites.
            cute.arch.fence_proxy("async.shared", space="cta")
            self.epilogue(args, ring, tmem, cta, n)
        else:
            self.multiply(ring, bars, tmem, n, stage)
        release_tmem(tmem, columns, warp, 4)

epilogue(args, ring, tmem, cta, n)

D rows 16 * warp + lane (lane < 16: M=64 TMEM layout) -> FP32 [token, row] tile in the drained ring -> 8-row BF16 chunks to the mailbox / partials.

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
@cute.jit
def epilogue(self, args, ring, tmem, cta, n: Constexpr):
    """D rows 16 * warp + lane (lane < 16: M=64 TMEM layout) -> FP32 [token, row]
    tile in the drained ring -> 8-row BF16 chunks to the mailbox / partials."""
    x, _, mailbox, mc, slot = args[:5]
    M: Constexpr = x.shape[0]
    tid, warp, lane, _ = thread()
    values = load_tmem(tmem + (32 * warp << 16), self.geometry(M)[3])
    tile = view(ring.iterator, 0, Float32, (n, 68), (68, 1))  # [token, row]
    if lane < 16:
        for t in cutlass.range_constexpr(M):
            tile[t, 16 * warp + lane] = values[t]
    named_barrier(1, 128)
    for i in cutlass.range_constexpr((M * 8 + 127) // 128):
        c = tid + 128 * i
        if c < M * 8:
            token, part = c // 8, c % 8
            reduced = cute.make_rmem_tensor(8, Float32)
            for h in cutlass.range_constexpr(2):
                words = smem_words(at(tile, token, part * 8 + h * 4))
                for v in cutlass.range_constexpr(4):
                    reduced[h * 4 + v] = bits_float(words[v])
            words = pack8(reduced, "latent")
            column = cta * 64 + part * 8
            if cutlass.const_expr(self.twoshot):
                push(mailbox, mc, slot, token, column, words)
            else:
                offset = cute.crd2idx((slot, token, column), mailbox.layout)
                multicast_words(mc + Int64(offset) * 2, words)

geometry(batches)

(N, ring stage bytes, stages, TMEM columns).

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
315
316
317
318
319
320
def geometry(self, batches):
    """(N, ring stage bytes, stages, TMEM columns)."""
    n = (batches + 7) // 8 * 8
    stage = 8192 + n * 128
    stages = min(self.blocks, self.budget // stage)
    return n, stage, stages, max(32, 1 << (n - 1).bit_length())

load(source, ring, rows, valid, row0, column)

cp.async rows x 64 BF16 of source from (row0, column) into a K-major 128B-swizzled tile; rows >= valid are left alone. valid < 0: weights (evict-first, every row).

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
@cute.jit
def load(self, source, ring, rows: Constexpr, valid: Constexpr, row0, column):
    """cp.async ``rows`` x 64 BF16 of ``source`` from (row0, column) into a K-major
    128B-swizzled tile; rows >= ``valid`` are left alone. ``valid`` < 0: weights
    (evict-first, every row)."""
    tid, _, _ = cute.arch.thread_idx()
    for i in cutlass.range_constexpr((rows * 8 + 127) // 128):
        c = tid + 128 * i
        row, chunk = c // 8, c % 8
        dst = ring + row // 8 * 1024 + row % 8 * 128 + (chunk ^ row % 8) * 16
        src = at(source, row0 + row, column + chunk * 8)
        if cutlass.const_expr(valid < 0):
            stage_weight(dst, src)
        elif cutlass.const_expr((i + 1) * 128 <= valid * 8):
            copy_g2s(dst, src)
        else:
            if c < valid * 8:  # noqa: SIM102
                copy_g2s(dst, src)

multiply(ring, bars, tmem, n, stage)

The MMA warp: wait for each K block, issue its four K16 MMAs from the elected lane and release its stage (the last block: the accumulator).

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
@cute.jit
def multiply(self, ring, bars, tmem, n: Constexpr, stage: Constexpr):
    """The MMA warp: wait for each K block, issue its four K16 MMAs from the
    elected lane and release its stage (the last block: the accumulator)."""
    stages: Constexpr = self.geometry(n)[2]
    # kind::f16 descriptor: FP32 D, BF16 A and B, both K-major, N, M = 64
    instruction: Constexpr = 1 << 4 | 1 << 7 | 1 << 10 | n >> 3 << 17 | 4 << 24
    for kb in cutlass.range_constexpr(self.blocks):
        s: Constexpr = kb % stages
        wait(bars.iterator + s, kb // stages % 2)
        cute.arch.fence_proxy("async.shared", space="cta")
        fence_tmem()
        a = smem_desc(ring.iterator + s * stage, True)
        b = smem_desc(ring.iterator + s * stage + 8192, True)
        last: Constexpr = kb + 1 == self.blocks
        release: Constexpr = last or kb + stages < self.blocks
        barrier = bars.iterator + (2 * stages if last else stages + s)
        mma_bf16_block(tmem, a, b, instruction, int(kb > 0), barrier, int(release))

pack8(values, policy='plain')

Four BF16x2 words of values[0:8].

Source code in vllm/models/kimi_k3/nvidia/specialized_decode/cutedsl/oproj/producer.py 
47
48
49
50
51
def pack8(values, policy="plain"):
    """Four BF16x2 words of ``values[0:8]``."""
    return tuple(
        pack_pair(values[2 * j], values[2 * j + 1], policy=policy) for j in range(4)
    )